Do elite users' data matter for frontier models? A debate on RL datamix signals
plausaible · x · 2026-09-09
A technical debate on the value of user data for frontier models. kalomaze argues process improvements from user data overwhelmingly can't be causally attributed to individual power users — instead, aggregate signals like domain-wide power-user frustration redirect effort into that domain.
plausaible partially agrees, especially on the 2026 model datamix, but contends the views aren't mutually exclusive: cutting-edge knowledge genuinely comes from a few select people worth farming. A representative argument over elite-data vs aggregate-signal strategies in RL data collection.
More from Models
- DeepSeek v4 and GLM Now Run Faster Than vLLM and SGLang — jedisct1 · 2026-09-09
- User claims running 'gpt-6-astra' for over an hour at just $7.06 — arthurcolle · 2026-09-09
- Qwen3.8 27B quantization benchmark: 4-bit holds up, 1-bit collapses — victormustar · 2026-09-09
- Dev claims running gpt-6-astra for over an hour cost just $7.06 with his own harness — arthurcolle · 2026-09-09
- Blogger argues OpenAI's latest result must be post-training, not pre-training — kimmonismus · 2026-09-09
- Debate: is pretraining size scaling dead, or just architecture scaling conflated? — samsja19 · 2026-09-09